A code-generation language model trained entirely on a smartphone.
GPT-2 architecture. Snapdragon 685. CPU only. 50+ hours. $0.00.
Yuuki (勇気 - courage/bravery) is an open-source code-generation language model being trained entirely on a Redmi 12 smartphone with a Snapdragon 685 processor running in CPU-only mode. No cloud GPUs. No training clusters. No budget. Just a phone, patience, and the determination to make AI accessible to everyone.
Started in January 2026 by agua_omg, a young independent developer who couldn't afford Claude subscriptions anymore, Yuuki represents a radical experiment in democratizing AI training. The project proves that you don't need expensive infrastructure to contribute to the AI ecosystem -- you just need time, creativity, and a willingness to think differently.
The name combines the Japanese word for snow (Yuki) with the character Yuu from Girls' Last Tour, reflecting both the project's Japanese aesthetic and its scrappy, survival-against-the-odds spirit.
|
Training Progress
|
Model Performance (Checkpoint 2000)
Scores improving with each checkpoint. |
|
Yuuki Chat Yuuki Web |
yuy yuy-chat |
| Repository | Description | Language |
|---|---|---|
| yuuki-training | Training code, scripts, and configuration for the Yuuki models | Python |
| yuy | CLI for model management and local inference | Rust |
| yuy-chat | TUI chat interface for terminal-based conversations | Rust |
| Yuuki-web | Official landing page and project showcase | TypeScript |
| Yuuki-chat | Web-based chat interface with research and YouTube modes | TypeScript |
Detailed technical documentation, training methodology, evaluation benchmarks, and the complete story of training a language model on a smartphone is available in our research paper:
Topics covered: hardware constraints, training optimization for mobile CPUs, dataset composition, evaluation methodology, checkpoint comparisons, and future roadmap.
Join the Yuuki community to discuss the project, share results, contribute ideas, and connect with other developers interested in democratizing AI training.
|
Accessibility First AI training shouldn't require a PhD, a lab, or a six-figure cloud bill. Yuuki proves that meaningful contributions to open-source AI can come from anyone with a consumer device and determination. Radical Transparency Every training step, every checkpoint, every mistake is documented and shared publicly. The goal isn't just to build a model -- it's to show others that they can do it too. |
Zero-Budget Innovation By embracing extreme hardware constraints, Yuuki explores training techniques that work on edge devices. These optimizations benefit everyone -- from hobbyists to enterprises looking to reduce costs. Community-Driven Yuuki belongs to the community. Whether you contribute code, share ideas, report bugs, or just star the repos -- you're part of this journey. |
| Category | Technologies |
|---|---|
| Model Training | Python, PyTorch, Transformers, GPT-2 architecture |
| CLI Tools | Rust, Tokio, Clap, Ratatui, HuggingFace Hub |
| Web Apps | Next.js 16, React 19, TypeScript, Tailwind CSS v4 |
| APIs | HuggingFace Inference API, Tavily Search, YouTube Data API v3 |
| Deployment | Vercel, HuggingFace Spaces |
| Design | Dark cyberpunk aesthetic, sakura pink (#f472b6) accents, Geist font |
Yuuki is a zero-budget project built by an independent developer. If you find value in this work and want to support continued development, training, and community building, consider sponsoring:
Every contribution helps keep the phone charged and the training running.
Yuuki started out of necessity. In January 2026, agua_omg -- a young developer -- could no longer afford Claude API subscriptions. But instead of giving up on AI entirely, they asked a different question: What if I could train my own model with nothing but the phone in my pocket?
The name "Yuuki" is a fusion of two inspirations:
- 雪 (Yuki) -- Japanese for "snow", representing purity and the cold reality of having zero budget
- ユー (Yuu) -- The character from Girls' Last Tour, an anime about survival in a post-apocalyptic world with minimal resources
Together, they form 勇気 (Yuuki) -- courage -- which perfectly captures the audacity of training a language model on a smartphone.
The training happens on a Redmi 12 with a Snapdragon 685 processor, running pure CPU inference with no GPU acceleration. Each training step takes seconds. Each epoch takes hours. The phone gets warm. The progress is slow. But it's working.
Yuuki is proof that innovation doesn't require money -- it requires persistence, creativity, and a willingness to try the impossible.
Built with patience, a phone, and zero budget.
Training the future of AI, one step at a time.